Abstract

In recent years, the industry has sought insights from abundant data generated by drilling operations. One of the key focus areas is the rate of penetration (ROP) which impacts costs directly, and emissions indirectly. Previous work has succeeded in predicting and optimizing ROP, however was limited to specific fields and small-scale applications. This limitation stems from unobserved information between different fields or operations that often impacts model usability. This paper provides a new way of well planning by leveraging the power of unsupervised machine learning to deliver higher drilling efficiency, lower costs, and less uncertainty.

Unsupervised machine learning techniques are used to infer information about drilling operations from real-time data that is not directly measured. Certain well types seem to be separable in low dimensions, based on qualitative interpretation of clusters visualized in 2D, and according to analyzing cluster membership based on which wells the data came from, and by associating common characteristics of wells within clusters using external information.

This work introduces a novel approach to collect, separate, and extract value from data that is otherwise unused. Data synthesized by Variational Autoencoders could be used for enhanced well planning, sold as a standalone product, or shared between industry players with reduced privacy concerns, increasing the wealth of data available.

This content is only available via PDF.
You do not currently have access to this content.